fx

index.html

@@ -160,7 +160,7 @@
     // Keep RAM low for mobile: small context + FP16 V‑cache (WASM safe)
     const LOAD_CONFIG = {
       n_ctx: 768,
-      n_batch: 48,
+      n_batch: 128,  // must be >= 64 to satisfy GGML_KQ_MASK_PAD and avoid batch overflow in wasm
       cache_type_k: "q4_0",   // int4 K cache: reduces RAM without flash_attn
       cache_type_v: "f16",    // IMPORTANT: V cache quant requires flash_attn; not available in WASM
       flash_attn: false,
@@ -267,7 +267,7 @@
       $input.value = '';
 
       const assistantBubble = ui.add('assistant', '');
-      truncateHistoryForMobile(900);
+      truncateHistoryForMobile(600); // trim harder to reduce initial prompt size vs 768 ctx
 
       $send.disabled = true; $stop.disabled = true; // will flip to true once stream starts
       aborter = new AbortController();